SLMs as Multi-Agent Routers: A Progressive SFT and Reinforcement Learning Approach
Quick Answer
This paper shows that A new approach using supervised fine-tuning and reinforcement learning trains a small language model for optimal agent selection in retrieval tasks, achieving an NDCG@10 of 0.918, significantly outperforming intent-based models like Amazon Nova Lite and Claude Haiku 4.5.
Quick Take
The model reduces selection latency by 82.4%, making it more efficient for query routing.
Key Points
- The model learns agent suitability based on retrieval performance and query distributions.
- Achieves a mean NDCG@10 of 0.771, outperforming Nova Lite by 0.177.
- Selection latency is reduced to 120.1ms, an 82.4% improvement over Nova Lite.
- Utilizes a hierarchical reward function for better agent-query alignment.
- Addresses limitations of intent-based routing by incorporating content relevance.
DeepSignal Analysis
What happened
A small language model was trained using supervised fine-tuning and reinforcement learning to improve agent selection for retrieval tasks. This model achieved an NDCG@10 score of 0.918, outperforming intent-based models like Amazon Nova Lite and Claude Haiku 4.5, which scored 0.539 and 0.490, respectively. Additionally, the model reduced selection latency by 82.4%, making it more efficient for query routing.
Key evidence
- The trained model achieved an NDCG@10 of 0.918, significantly higher than the 0.539 and 0.490 scores of Amazon Nova Lite and Claude Haiku 4.5, respectively.
- The model reduced selection latency to 120.1ms, representing an 82.4% decrease compared to Nova Lite.
- The approach combines supervised fine-tuning with reinforcement learning to enhance agent selection and parameter generation based on retrieval relevance.
Why it matters
This advancement in agent selection for retrieval tasks highlights the limitations of intent-based models, which do not account for the relevance of retrieved content. By incorporating a hierarchical reward function, the new model can better identify suitable agents for specific queries, potentially leading to more accurate and efficient information retrieval systems. The significant reduction in selection latency also suggests practical benefits for real-time applications.
Paper Resources
📖 Reader Mode
~2 min readAbstract:Specialised retrieval agents typically surface higher quality results than general-purpose search, but selecting the optimal agent for a given query remains an open problem. Current approaches route queries based on inferred topic or intent, however intent-based selection is fundamentally limited: it does not incorporate signal from retrieved content, and cannot detect when a topically aligned agent produces low-relevance results. We address this by training a small language model via supervised fine-tuning followed by reinforcement learning to jointly perform agent selection and structured parameter generation for downstream tool calls, using a hierarchical reward function grounded in retrieval relevance along with query-agent topic alignment. This enables the model to learn task-dependent agent suitability from retrieval performance: which agents reliably yield high-relevance results for which query distributions, and when to redirect queries away from specialised agents despite surface-level topical overlap. On a targeted subset of such agent-query mismatches, the trained model achieves an NDCG@10 of 0.918 compared to 0.539 and 0.490 for two LLM baselines (Amazon Nova Lite and Claude Haiku 4.5) that route on intent alone. Overall, it achieves a mean NDCG@10 of 0.771 (+0.177 over Nova Lite, +0.219 over Haiku) with a mean selection latency of 120.1ms, an 82.4% reduction over Nova Lite.
| Subjects: | Computation and Language (cs.CL); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2608.00030 [cs.CL] |
| (or arXiv:2608.00030v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2608.00030 arXiv-issued DOI via DataCite |
Submission history
From: Venkatashesha Gayathri Kondapalli [view email]
[v1]
Wed, 15 Jul 2026 11:11:39 UTC (129 KB)
— Originally published at arxiv.org
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from arXiv cs.CL
See more →TriAgent: Divergence-Aware Committees for Cost-Efficient Financial Sentiment Analysis
TriAgent introduces a cost-efficient multi-agent system for financial sentiment analysis, combining VADER, FinBERT, and Qwen2.5. It achieves an F1 score of ~0.87 with significant savings of $9.3M/year at a 10M-user scale compared to GPT-4o-mini, while also detecting hallucinations with an AUC of 0.90.